Is there any good practice related to dynamic_cast error handling (except not using it when you don't have to)? I'm wondering how should I go about NULL and bad_cast it can throw. Should I check for both? And if I catch bad_cast or detect NULL I probably can't recover anyway... For now, I'm using assert to check if dynamic_cast returned not NULL value. Would you accept this solution on a code review?
If the dynamic_cast should succeed, it would be good practice to use boost::polymorphic_downcast instead, which goes a little something like this:
assert(dynamic_cast<T*>(o) == static_cast<T*>(o));
return static_cast<T*>(o);
This way, you will detect errors in the debug build while at the same time avoiding the runtime overhead in a release build.
If you suspect the cast might fail and you want to detect it, use dynamic_cast and cast to a reference type. This cast will throw bad_cast in case of error, and will take down your program. (This is good if, as you say, you are not going to recover anyway)
T& t = dynamic_cast<T&>(o);
t.func(); //< Use t here, no extra check required
Use dynamic_cast to a pointer type only if the 0-pointer makes sense in the context. You might want to use it in an if like this:
if (T* t = dynamic_cast<T*>(o)) {
t->func(); //< Use t here, it is valid
}
// consider having an else-clause
With this last option you need to make sure that the execution path makes sense if the dynamic_cast returns 0.
To answer your question directly: I would prefer one of the two first alternatives I have given to having an explicit assert in the code :)